TOPS: The AI jargon that even big tech is sick of

Intel Lunar Lake chips at Computex 2024
Pictured: Intel Lunar Lake processor at Computex 2024
// Let's talk TOPS.
Fergus Halliday
Jun 07, 2024
Icon Time To Read3 min read

I don’t know if 2024 will go down as the yearof AI but it certainly feels like the year it became impossible to escape it.

CES 2024 was dominated by AI tech and Computex 2024 didn’t deviate from that trend. Taipei’s biggest technology tradeshow was filled with folks looking to shill their AI-powered wares like a hot dog vendor at a baseball game.

If you’re looking to get your head around the many new laptops (which are being billed as AI PCs) announced at the tradeshow, TOPS is a key term you’ll want to understand. Manufacturers have moved towards adopting the term earlier this year and started using it as a go-to benchmark for AI PC and NPU performance.

For those unfamiliar, an NPU is a neural processing unit. It’s a specific part of a PC that’s designed to handle AI-related workloads in the same way that a GPU is built to handle stuff like games and 3D graphics. While most older PCs don’t have a built-in NPU, many of the latest AMD, Intel and Qualcomm processors do feature one.

Some NPUs are better than others but assuming that they outlast this current wave of hype around AI technologies, the PC market will eventually reach a place where their presence becomes as ubiquitous as a CPU or GPU. Given that, the conversation has begun to shift from whether or not a PC has an NPU to what that hardware can do. As mentioned above, not all NPUs are born equal.

Enter TOPS. This is an easy shorthand for Trillions of Operations Per Second or tera operations per second. The higher it is, the faster a given PC can process AI-related tasks and applications. The faster a PC does anything, the more time you’re “saving” by using it. Even if AI is being marketed as the next big thing, the underlying logic here isn’t nearly so much as a departure from traditional PC speak than the marketing might have you believe.

TOPS is also being used as a way to distinguish between different sub-segments of the AI PC landscape.

While most PCs are capable of running AI applications without an NPU, an AI PC with dedicated AI hardware can run them much better. Hence, the rise of the term AI PCs. You’ve also got Microsoft’s new Copilot Plus PC branding, which covers any AI PC with at least 40 TOPS of performance.

Even if all this largely boils down to marketing, there’s still some utility here for consumers. Understanding how the brands looking to sell you your next PC are thinking about the category writ large helps you to understand what features and performance will suit your needs.

Of course, the thing about anything that’s measured in trillions is that most people aren’t equipped to grapple with numbers of that magnitude, let alone ones that relate to computational workloads. That’s literally the problem that computers exist to solve but while the technology industry is already talking TOPS, you have to wonder whether the term means anything to those who aren’t PC enthusiasts or AI advocates.

Speaking to, vice president and GM for Intel’s client AI and technical marketing division Robert Hallock laid out his rationale for TOPS as an AI metric in fairly digestible terms.

“Every device needs a spec,” he explained.

“For better or for worse, TOPS has become that. It is the measure of possible performance and the GPU market went through this some years ago. We all remember gigaflops or teraflops, right?”
Intel AI PC Header

Hallock noted that this term for GPU performance doesn’t really get used anymore and has gradually fallen out of favour relative to benchmarks based on practical testing of features and real-life workloads.

“I personally — or even professionally — do not like TOPS as a measurement. It doesn't tell anyone anything about the experience.”

According to him, “the optimization that's inside and the software stack is very complicated for AI workloads and so we want to move to a place where we get to get to something that feels a little more familiar [for] performance benchmarking.”

Hallock predicted that in the future you’ll be able to go online and see benchmarks for AI workloads just like you can gaming ones today. He said that Intel is already working with several benchmark vendors to make it easier to produce quality performance analytics for AI workloads and models.

Over the next few years, Hallock predicted that AI performance would follow a similar trajectory to that of GPU performance as a result of the natural similarities between the two components.

“Graphics, as we all know, has been on a steady climb upwards in performance for a very long time and you should expect that AI will do the same,” he said.

In parallel with this trend, Hallock suggested that techniques like quantizing would help reduce AI workload requirements over time and would fuel further performance gains for consumers. In the long run, he echoed Intel’s perspective that AI performance will become not just an aspect of the PC market but inextricably tied to it.

“We think that by 2028 about 80 per cent of the PCs will look like [that]. They'll have extensive AI acceleration inside and most applications will be based on it”

“The best analogy I can give you is graphics cards [where] 20 years ago people were like ‘why did you do that?’”

Disclosure: Australia's coverage of Computex 2024 is supported by MSI.

Fergus Halliday
Written by
Fergus Halliday
Fergus Halliday is a journalist and editor for He’s written about technology, telecommunications, gaming and more for over a decade. He got his start writing in high school and began his full-time career as the Editor of PC World Australia. Fergus has made the MCV 30 Under 30 list, been a finalist for seven categories at the IT Journalism Awards and won Most Controversial Writer at the 2022 Consensus Awards. He has been published in Gizmodo, Kotaku, GamesHub, Press Start, Screen Rant, Superjump, Nestegg and more.

Related Articles